CN111832839B - Energy consumption prediction method based on sufficient incremental learning - Google Patents

Energy consumption prediction method based on sufficient incremental learning Download PDF

Info

Publication number
CN111832839B
CN111832839B CN202010723405.3A CN202010723405A CN111832839B CN 111832839 B CN111832839 B CN 111832839B CN 202010723405 A CN202010723405 A CN 202010723405A CN 111832839 B CN111832839 B CN 111832839B
Authority
CN
China
Prior art keywords
data
value
incremental learning
energy consumption
similarity value
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202010723405.3A
Other languages
Chinese (zh)
Other versions
CN111832839A (en
Inventor
刘晶
于兵
董瑶
赵佳
金玉蓉
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hebei University of Technology
Original Assignee
Hebei University of Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Hebei University of Technology filed Critical Hebei University of Technology
Priority to CN202010723405.3A priority Critical patent/CN111832839B/en
Publication of CN111832839A publication Critical patent/CN111832839A/en
Application granted granted Critical
Publication of CN111832839B publication Critical patent/CN111832839B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G06N20/10—Machine learning using kernel methods, e.g. support vector machines [SVM]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
    • G06Q10/063—Operations research, analysis or management
    • G06Q10/0639—Performance analysis of employees; Performance analysis of enterprise or organisation operations
    • G06Q10/06393—Score-carding, benchmarking or key performance indicator [KPI] analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/10—Services
    • G06Q50/26—Government or public services
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
    • Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
    • Y02P90/80—Management or planning
    • Y02P90/82—Energy audits or management systems therefor

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Human Resources & Organizations (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Economics (AREA)
  • Strategic Management (AREA)
  • General Physics & Mathematics (AREA)
  • Tourism & Hospitality (AREA)
  • Development Economics (AREA)
  • Educational Administration (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Software Systems (AREA)
  • General Business, Economics & Management (AREA)
  • Marketing (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • General Health & Medical Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Health & Medical Sciences (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Game Theory and Decision Science (AREA)
  • Mathematical Physics (AREA)
  • Medical Informatics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Molecular Biology (AREA)
  • Primary Health Care (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses an energy consumption prediction method based on sufficient incremental learning, which comprises the following contents: sorting the multidimensional parameters of the energy consumption data, extracting the models of different hubs, and using the models to perform effective feature extraction on the sample by a weighted kernel principal component analysis method; then grouping the data, calculating and judging similarity values in the process of each round of incremental learning, performing multi-round incremental learning according to different degrees of the similarity values, and continuously and dynamically adjusting and judging threshold values, data samples and weighted values; and finally, using an error back propagation neural network to predict energy consumption, and using a Root Mean Square Error (RMSE) value to evaluate a prediction result. The invention verifies the effectiveness of the prediction method, and the finally obtained RMSE value is as low as 0.000112, and the effect is optimal compared with other comparison methods.

Description

Energy consumption prediction method based on sufficient incremental learning
Technical Field
The invention relates to the technical field of energy consumption prediction, in particular to an energy consumption prediction method based on sufficient incremental learning.
Background
Energy consumption and environmental pollution are a serious examination accompanying the sustainable and high-speed development of economy in China, wherein industrial energy consumption accounts for nearly 70% of the total energy consumption of the society in China, so that in the industrial field, the improvement of energy use efficiency and the reduction of energy consumption are key points for accelerating the promotion of green transformation of industry. With the civilization of vehicles, vehicle accessories are also more and more concerned by the public, and in order to meet the requirements of customers on the personalized customization of the vehicle accessories, and the wheel hub as an important component part of the vehicle accessories has decorative appearance, the personalized wheel hub is also more and more popular. However, when the hub is produced, the use of energy consumption cannot be effectively controlled, and the energy consumption data can influence the cost for producing the hub, so that the method has important significance for researching the energy consumption of the hub product.
In the industrial field, the use of energy consumption is getting more attention. In the revolution of industry 4.0, the integration of advanced technologies such as the internet of things, artificial intelligence and industrial internet, and the application of methods such as data acquisition, data management, analysis modeling and the like in the industrial production process provide support for the prediction of energy demand. At present, time series prediction, a neural network, gray prediction, an extreme learning machine, a support vector machine and other methods are mainly adopted for prediction, and good research results are obtained. An article [ Wangzuo and the like, building energy consumption prediction [ J ] of a hybrid model constructed by improving a whale algorithm, computer measurement and control 2020,28(02):197 + 201+205 ] proposes a novel hybrid model for short-term prediction of building energy consumption, processes data through a complementary set empirical mode decomposition method and a feedback neural network optimized by an improved whale optimization algorithm to obtain a final energy consumption prediction value, and has good prediction performance. In the article [ Zhousheason et al, incremental ELM virtual machine energy consumption prediction [ J ] automated chemical newspaper, 2019,45(07):1290-1297 ] based on the compressed momentum term, aiming at the problem that a prediction model based on the traditional incremental extreme learning machine has redundant nodes, the compressed momentum term is added to feed back network training errors to the output of a hidden layer, so that the prediction result is closer to an output sample, the redundant hidden layer nodes are reduced, and the convergence speed is accelerated. An article [ Liujia school and the like, an airline energy consumption sequence prediction [ J ] based on LSTM, computer application and software, 2019,36(10):60-65 ] proposes a long-short time memory network prediction model based on grid search algorithm optimization, and improves the accuracy of energy consumption prediction. An article [ Yangjingjun. regional building electric power energy consumption rapid prediction simulation [ J ] computer simulation, 2019,36(04): 432-. An article [ Chen Jing Jie et al, unsaturated airport energy consumption prediction [ J ] based on a two-step decomposition method and SARIMA, computer application and software, 2019,36(04):46-50+78 ] proposes a combined prediction method based on a two-step decomposition method and a seasonal differential autoregressive moving average model, combines an adaptive noise and variational modal decomposition method, and adopts the seasonal differential autoregressive moving average model to model and predict data, thereby greatly improving the accuracy of energy consumption prediction. An article [ GuoJun, energy consumption prediction of polycrystalline silicon production based on LSTM-Adaboost [ J ] computer application and software, 2018,35(12):71-75+117 ], provides an energy consumption prediction model based on the LSTM-Adaboost recurrent neural network polycrystalline silicon production process, optimizes an LSTM objective function by using a regularization method and introduces an Adaboost algorithm to optimize the LSTM model by combining a PCA algorithm to obtain the prediction model, and realizes effective energy consumption prediction. An article [ Dingfeihong and the like, a short-term building energy consumption prediction model [ J ] based on a genetic optimization decision tree, computer engineering, 2019,45(06):280 plus 289+296 ] proposes a genetic optimization decision tree model, the subtree generation process of the decision tree is promoted by optimizing the gradient through a genetic algorithm, and higher prediction precision is obtained. An article [ high school fund and the like ] LS-SVM subway station air conditioning system energy consumption prediction model [ J ] based on ISOA, computer and modernization, 2018(10):36-43 ] provides a method for optimizing model parameters in an LS-SVM modeling process by using a crowd search algorithm improved from two aspects of algorithm search step length and search direction, and the prediction precision and speed are improved. An article [ Wangkun et al, LSSVM for airport energy consumption prediction [ J ] based on EMD and fruit fly parameter optimization [ computer age, 2017(04):35-40 ] proposes an energy consumption prediction method of a least square support vector machine by combining empirical mode decomposition and fruit fly parameter optimization, and improves the prediction precision.
It can be seen from the above documents that, in the process of performing energy consumption prediction, although methods such as dimension reduction and incremental learning are used, information of the attributes of data itself is not fully utilized, and in the process of incremental learning, only the problem of redundant nodes is solved, but for increasing data, effective data is selected to perform efficient energy consumption prediction, which is a problem that needs to be solved urgently at present. The invention improves the use of the label in the dimension reduction process, and uses the attribute of different types of the hub as the label in the weight, thereby realizing the high-efficiency utilization of the data information; in the incremental learning process, on one hand, the weight of effective data is improved, on the other hand, the weight of inefficient data is reduced, and the data is not fully utilized after one-time screening is finished, so that multi-layer screening incremental learning is performed in a circulating mode, all effective data are finally left, and more accurate energy consumption prediction is finally realized.
Disclosure of Invention
The invention aims to:
the invention provides an energy consumption prediction method based on sufficient incremental learning, which is characterized by combining the types of hubs and using a weighted kernel principal component analysis method to extract effective characteristics of a sample, then calculating and judging similarity values, carrying out multi-round incremental learning according to different degrees of the similarity values, continuously and dynamically adjusting judgment threshold values, and finally using an error back propagation model to carry out energy consumption prediction. Compared with the original method, the method has the advantages that effective information is reserved to the maximum extent through characteristic extraction of data, the important degree relation included among energy consumption data is effectively solved through multiple rounds of incremental learning, dynamic adjustment of the energy consumption data is achieved, the effect of efficiently utilizing data information is achieved, and accurate energy consumption prediction is finally achieved.
The technical scheme adopted by the invention is as follows:
an energy consumption prediction method based on sufficient incremental learning comprises the following steps:
(1) sorting the multidimensional parameters of the energy consumption data, and extracting the models of different hubs;
(2) reducing the dimension of the sorted energy consumption data by adopting a weighted kernel principal component analysis method;
(3) dividing the data subjected to dimensionality reduction into training data and test data, grouping the training data and endowing the grouped training data with an initial weight value;
(4) performing first round learning of incremental learning on the first group of training samples and the weighted values, and calculating to obtain an average similarity value of the first group of training data;
(5) continuously inputting the residual data into an incremental learning model, and respectively comparing the residual data with the first group of training data to judge similarity values to perform different operations;
(6) when the similarity between the samples is greater than the average similarity value, increasing the weight value of the existing samples; when the similarity value between the samples is smaller than the average similarity value, reducing the weight value of the existing samples, uniformly storing the samples together for the incremental learning of the next round, and realizing sufficient incremental learning through multiple rounds of incremental learning;
(7) in the process of the operation, the threshold parameter, the data sample and the weight value of each round of incremental learning are adjusted along with the difference of data;
(8) when the average similarity value of the data used for incremental learning at a time is smaller than 1/2 of the initial average similarity value, the incremental learning is finished, and at the moment, the data with the ownership weight value smaller than the initial weight value 1/2 is deleted;
(9) inputting the obtained data into an error back propagation model for energy consumption prediction;
(10) and evaluating the energy consumption prediction result by adopting an evaluation index RMSE value.
The weighting of the weighted kernel principal component analysis method in the step (2) is calculated by processing labels of samples with different hub models as weights, constructing a weighted kernel matrix by fusing information values of the labels in the kernel matrix, wherein the weights corresponding to data in the weighted kernel matrix are
Figure BDA0002600827230000041
Wherein u isiAnd representing the weight vector corresponding to the ith sample.
In the step (5), the Euclidean distance is used for calculating the similarity value of the multi-dimensional data, and the similarity value calculation formula is as follows:
Figure BDA0002600827230000042
where dist (x, y) represents the Euclidean distance between two different failure features, namely the new feature x and the existing feature y (the existing feature contains the initial feature and the new feature for completing the new operation), that is
Figure BDA0002600827230000043
Wherein x isiAnd yiThe ith value representing the new feature x and the existing featureThe ith value of y, the difference between x and y is smaller, and the Euclidean distance is smaller, the similarity value is larger, so that the similarity between the newly added feature and the existing feature is measured.
The process of sufficient increment learning in the step (6) comprises the following steps:
in the whole process of sufficient incremental learning, firstly giving initial weight values to all data, secondly calculating similarity values among the first group of data, calculating to obtain an average similarity value, then respectively calculating the similarity values of the other newly added data and the first group of data, and increasing the weight values of the existing samples when the similarity values are larger than the existing average similarity value; and when the current average similarity value is smaller than the existing average similarity value, reducing the weight value of the existing sample and storing all data together for carrying out the increment learning of the next round. In the next round of incremental learning, firstly grouping the existing data, secondly calculating the similarity value between the first group of data, calculating to obtain an average similarity value, then respectively calculating the similarity values of the other newly added data and the first group of data, and increasing the weight value of the existing sample when the similarity value is greater than the average similarity value; when the average similarity value is smaller than the average similarity value, the weight value of the existing sample is reduced, all data are stored together for incremental learning in the next round, until 1/2 that the average similarity value of the data for incremental learning is smaller than the initial average similarity value at a certain time, the incremental learning is finished, at this time, the data with the ownership weight value smaller than the initial weight value 1/2 are deleted, and the whole process of sufficient incremental learning is completed.
In order to solve the problem of energy consumption of a hub in the production process, the patent provides an energy consumption prediction method based on sufficient incremental learning, a weighting kernel principal component analysis method is used for extracting effective characteristics of a sample, similarity values are calculated and judged, the weights of an initial data sample and a newly added data sample are dynamically adjusted by using the sufficient incremental learning according to different conditions of the similarity values, and finally an error back propagation model is used for predicting the energy consumption, so that the efficient incremental learning and the accurate energy consumption prediction are realized.
Firstly, the theoretical basis of the method of the invention is as follows:
1. weighted kernel principal component analysis: although the kernel principal component analysis method can process the nonlinear problem, the information in the data label is ignored, so the invention adopts the weighted kernel principal component analysis method (WKPCA) to reduce the dimension of the energy consumption data, and retains effective information as much as possible, thereby improving the efficiency of energy consumption prediction and reducing the error of the energy consumption prediction value.
2. And (3) sufficient increment learning: in the whole process of sufficient incremental learning, firstly giving initial weight values to all data, secondly calculating similarity values among the first group of data, calculating to obtain an average similarity value, then respectively calculating the similarity values of the other newly added data and the first group of data, and increasing the weight values of the existing samples when the similarity values are larger than the existing average similarity value; and when the current average similarity value is smaller than the existing average similarity value, reducing the weight value of the existing sample and storing all data together for carrying out the increment learning of the next round. In the next round of incremental learning, firstly grouping the existing data, secondly calculating the similarity value between the first group of data, calculating to obtain an average similarity value, then respectively calculating the similarity values of the other newly added data and the first group of data, and increasing the weight value of the existing sample when the similarity value is greater than the average similarity value; when the average similarity value is smaller than the average similarity value, the weight value of the existing sample is reduced, all data are stored together for incremental learning in the next round, until 1/2 that the average similarity value of the data for incremental learning is smaller than the initial average similarity value at a certain time, the incremental learning is finished, at this time, the data with the ownership weight value smaller than the initial weight value 1/2 are deleted, and the whole process of sufficient incremental learning is completed.
3. BP neural network (BPNN): and (3) an error back propagation algorithm model is used for realizing accurate prediction of energy consumption through training data.
Secondly, designing an energy consumption prediction method based on sufficient incremental learning:
the energy consumption prediction method based on the sufficient incremental learning has the following working principle: performing feature extraction on the data after the multidimensional parameters are sorted by adopting a weighted kernel principal component analysis method, so that the data becomes more efficient, and meanwhile, the data processing efficiency is improved; performing multiple rounds of incremental learning on the extracted features, continuously learning new knowledge on the basis of keeping original data information, realizing sufficient incremental learning, and continuously and dynamically adjusting and judging a threshold value, a data sample and a weight value in the incremental learning process; and finally, inputting the data obtained by training into an error back propagation algorithm to accurately predict the energy consumption, wherein the operation of the method is shown in figure 1.
1. Energy consumption prediction method structure based on sufficient incremental learning
According to the weighted kernel principal component analysis method, the dimensionality of the data sample is reduced through feature extraction, and the data processing efficiency is further improved. And performing multi-round incremental learning on the obtained data, dynamically adjusting and judging a threshold value, a data sample and a weight value by calculating a similarity value, and inputting the obtained data with the largest information content into an error back propagation algorithm for energy consumption prediction.
2. Specific implementation of the Algorithm
The steps of the energy consumption prediction method based on the sufficient incremental learning are described as follows, and the flow chart is shown in FIG. 2.
(1) Sorting the multidimensional parameters of the energy consumption data, and extracting the models of different hubs;
(2) reducing the dimension of the sorted energy consumption data by adopting a weighted kernel principal component analysis method;
(3) dividing the data subjected to dimensionality reduction into training data and test data, grouping the training data and endowing the grouped training data with an initial weight value;
(4) performing first round learning of incremental learning on the first group of training samples and the weighted values, and calculating to obtain an average similarity value of the first group of training data;
(5) continuously inputting the residual data into an incremental learning model, and respectively comparing the residual data with the first group of training data to judge similarity values to perform different operations;
(6) when the similarity between the samples is greater than the average similarity value, increasing the weight value of the existing samples; when the similarity value between the samples is smaller than the average similarity value, reducing the weight value of the existing samples, uniformly storing the samples together for the incremental learning of the next round, and realizing sufficient incremental learning through multiple rounds of incremental learning;
(7) in the process of the operation, the threshold parameter, the data sample and the weight value of each round of incremental learning are adjusted along with the difference of data;
(8) when the average similarity value of the data used for incremental learning at a time is smaller than 1/2 of the initial average similarity value, the incremental learning is finished, and at the moment, the data with the ownership weight value smaller than the initial weight value 1/2 is deleted;
(9) inputting the obtained data into a BPNN for energy consumption prediction;
(10) and evaluating the energy consumption prediction result by adopting an evaluation index Root Mean Square Error (RMSE) value.
Adopt the produced beneficial effect of above-mentioned technical scheme to lie in:
the invention provides an energy consumption prediction method based on sufficient incremental learning, which is used for processing three processes of energy consumption data characteristic extraction, incremental learning and energy consumption prediction. The model firstly reserves effective data samples with label information by using a weighted kernel principal component analysis method, and then predicts the energy consumption after performing multiple rounds of incremental learning on the data by calculating the similarity value and according to the judgment condition, thereby realizing efficient incremental learning and accurate energy consumption prediction. Through experimental analysis, a weighted kernel principal component analysis method, sufficient incremental learning and an error back propagation model are combined together, the RMSE value of energy consumption prediction is 0.000112, and meanwhile, through experimental comparison, the method disclosed by the invention is superior to a long-short term memory network (LSTM), a Support Vector Regression (SVR) and an original error back propagation model (BP), and accurate energy consumption prediction can be realized.
Drawings
FIG. 1 is a block diagram of the present invention energy consumption prediction method based on full incremental learning;
FIG. 2 is a flow chart of a method for energy consumption prediction based on full incremental learning of the present invention;
FIG. 3 is a graph of training data set RMSE values and test data set RMSE values for a BP algorithm with an intermediate layer output dimension of 7 in the method of the present invention;
FIG. 4 is a graph of the training data set RMSE values and the test data set RMSE values for the BP algorithm for a middle layer output dimension of 4 in the method of the present invention;
FIG. 5 shows the RMSE values for the training data set and the RMSE values for the test data set for a selected number of iterations of 3000 in the method of the present invention;
Detailed Description
In order to better understand the technical solution of the present invention, the following detailed description is made with reference to the accompanying drawings and examples.
In the embodiment of the invention, an energy consumption prediction method based on sufficient incremental learning comprises the following steps:
(1) taking prediction of power consumption of a single hub product as an example, selecting 20 hub models for experiment, simulating partial data by Python due to less available data amount, obtaining related data from 1 month in 2015 to 12 months in 2018, sorting multi-dimensional parameters of all energy consumption data, and extracting models of different hubs;
(2) reducing the dimension of the sorted energy consumption data by adopting a weighted kernel principal component analysis method;
although the kernel principal component analysis method can process the nonlinear problem, the information in the data label is ignored, so the invention adopts the weighted kernel principal component analysis method (WKPCA) to reduce the dimension of the energy consumption data, and retains effective information as much as possible, thereby reducing the error of the predicted value. Assume a high-dimensional dataset X ═ X1,…,xi,…,xN]，xiRepresents the ith sample of the data set X, N represents the number of samples, and c represents the sample class. Let the weight vector of the sample be [ u ]1,…,ui,…,uN]Wherein u isiDenotes the ith sample xiAnd (4) corresponding weight vectors. Let the label information value of the sample be { l1,…,lj,…,lcIn which ljIndicating the jth tag information value. Sample xiIs determined by class information, i.e. sample xiWhen the group belongs to the j-th class,
ui＝lj
wherein, i is 1,2, …, N, j is 1,2, …, c. Since the Gaussian kernel function is adopted, the kernel function
Figure BDA0002600827230000091
Wherein, K (x)i,xj) The kernel matrix of i rows and j columns is represented, and sigma represents the width parameter of the function, and controls the radial action range of the function. Constructing a weighted kernel matrix according to the weight information of the sample
Figure BDA0002600827230000092
Figure BDA0002600827230000093
Wherein, Kw(xi,xj) A weighted kernel matrix, w, representing i rows and j columnsijIs the weight value of the ith row and the jth column in the weight matrix, and
Figure BDA0002600827230000094
constructing a weight matrix W ═ Wij}。
Thus obtaining Kw(xi,xj) The following were used:
Figure BDA0002600827230000095
according to the above formula, when ui＝ujTime, sample xiAnd sample xjBelong to the same model, wijWhen the weighted kernel matrix is equal to 1, the original kernel matrix K (x)i,xj) (ii) a When u isi≠ujTime, sample xiAnd sample xjBelong to different models, in this case wijIs a positive number greater than 0, and the class information between samples is introduced into the weighted kernel matrix Kw(xi,xj) In (1).
K is calculated by using the Jacobi methodwCharacteristic value λ of1,…,λi,…,λNAnd a feature vector alpha1,…,αi,…,αNWherein λ isiDenotes the ith characteristic value, αiRepresents the ith feature vector, i ═ 1,2, …, N; secondly, the characteristic value lambda is measurediPerforming descending arrangement and adjusting the corresponding characteristic vector alphai(ii) a Then orthogonalizing the eigenvector by Gram-Schmidt orthogonal method unit to obtain an orthogonal eigenvector betaiWherein beta isiRepresenting the ith orthogonal eigenvector. Calculating the cumulative contribution rate B of the obtained characteristic value1,…,Bi,…,BnIn which B isiRepresenting the ith cumulative contribution rate, i 1,2, …, n, given a contribution rate value p, if BtIf the number is more than or equal to p, extracting t main components beta1,…,βi,…,βtFinally, calculating the projection Y of the normalized sample X on the extracted feature vector, wherein Y is equal to Kw(i, j) · β, wherein β ═ β (β) is obtained1,…,βi,…,βt) And the obtained projection Y is data obtained by extracting WKPCA characteristics from the original data.
(3) Dividing the data subjected to dimensionality reduction into training data and test data, grouping the training data and endowing the grouped training data with an initial weight value;
(4) performing first round learning of incremental learning on the first group of training samples and the weighted values, and calculating to obtain an average similarity value of the first group of training data;
(5) continuously inputting the residual data into an incremental learning model, and respectively comparing the residual data with the first group of training data to judge similarity values to perform different operations;
when calculating the similarity between the new feature and the existing feature, the Euclidean distance is adopted to calculate the similarity Sim (x, y) of the multi-dimensional data
Figure BDA0002600827230000101
Where dist (x, y) represents the Euclidean distance between the newly added feature x and the existing feature y
Figure BDA0002600827230000102
Wherein x isiAnd yiRepresenting newly added features xThe difference between the ith value and the ith value of the existing feature y is smaller, the Euclidean distance is smaller, the similarity value is larger, and the similarity between the newly added feature and the existing feature is measured.
(6) When the similarity between the samples is larger than the average similarity value, increasing the weight value of the samples; when the similarity value between the samples is smaller than the average similarity value, reducing the weight value of the corresponding samples, uniformly storing the weight values together for the incremental learning of the next round, and realizing sufficient incremental learning through multiple rounds of incremental learning;
calculating the similarity between the newly added feature and the existing feature by using the formula in the step (5), selecting the maximum value Sim (x, y) as the similarity value of the feature, and judging how to process the weight of the feature according to the following principle:
using aiRepresenting the similarity value between two pairs of bit features in the ith round of increment learning, and the value of the similarity value is aiAver Sim (x, y), representing the average of the similarity between two features x and y in the existing features:
(1) if aiIf < Sim (x, y), then the new feature is highly similar to the existing feature, so the weight value of the existing feature is increased.
(2) If Sim (x, y) < aiIf the similarity between the new feature and the existing feature is lower than the average similarity, the similarity between the new feature and the existing feature is not high enough, so that the data meeting the condition are stored together, the weight value of the corresponding existing feature is correspondingly reduced, and then the next round of incremental learning is performed.
And calculating the dynamic weight of the feature according to the newly added feature increment principle to measure the change degree of the importance of the feature along with the change of time. Because the similarity of the newly added feature and the existing feature can reflect the importance of the newly added feature to the current model to a certain extent, the similarity value is normalized by using the following formula, and the dynamic weight of the similarity value is calculated.
Figure BDA0002600827230000111
Wherein, Sim (x, y)normAnd representing the similarity value of the newly added feature x and the existing feature y after normalization processing, wherein minSim (x, y) represents the minimum value of the similarity between the newly added feature x and the existing feature y, and maxSim (x, y) represents the maximum value of the similarity between the newly added feature x and the existing feature y.
For a feature that satisfies the condition in principle (1), since it has a higher similarity value, its weight is increased to enhance and remember the valid feature. The dynamic weight calculation method is as follows.
Wi+1＝Wi+Sim(x,y)norm
Wherein, Wi+1Weight value, W, representing the i +1 th improvement of the existing feature yiIndicates the ith increase of the weight value of the existing feature y, and WiAdding the similarity values of the newly added feature x and the existing feature y to obtain Wi+1When the first combination is carried out, i is 0, W0Is the initial weight value.
For the features satisfying the conditions in the principle (2), the similarity is low, so that the weight value is reduced, and after the first round of incremental learning is finished, the features are used for the second round of incremental learning.
Wi+1＝Wi-Sim(x,y)norm
Wherein, Wi+1Weight value, W, representing the i +1 th reduction of the existing feature yiRepresents the ith decrease of the weight value of the existing feature y, and WiObtaining W by making a difference with the similarity value of the newly added feature x and the existing feature yi+1When the weight value of the newly added feature x is reduced for the first time, i is 0, W0Is the initial weight value.
(7) In the process of the operation, the threshold parameter, the data sample and the weight value of each round of incremental learning are adjusted along with the difference of data;
(8) when the average similarity value of the data used for incremental learning at a time is smaller than 1/2 of the initial average similarity value, the incremental learning is finished, and at the moment, the data with the ownership weight value smaller than the initial weight value 1/2 is deleted;
(9) inputting the obtained data into a BPNN for energy consumption prediction;
(10) and evaluating the energy consumption prediction result by adopting an evaluation index Root Mean Square Error (RMSE) value.
The test verification of the energy consumption prediction method based on the sufficient incremental learning comprises the following steps:
1. description of data
The experimental data are derived from the hub characteristic parameters and energy consumption data of a hub manufacturing company, and the hub manufacturing company is provided with intelligent instruments in each production line for obtaining electric energy, water and natural gas consumed in the production process. In the experiment, the power consumption of a single hub is predicted as an example, 20 hub models are selected for carrying out the experiment, and as the available data amount is small, part of data is simulated by Python, and related data from 1 month in 2015 to 12 months in 2018 are obtained. As shown in table 1 below:
TABLE 1 characterization parameter description
Figure BDA0002600827230000121
2. Procedure of experiment
For energy consumption data of 20 wheel hub models, firstly, parameters of 'single model' are extracted to serve as labels of the data, then, dimension reduction is carried out on the remaining 10 parameters by using a weighted kernel principal component analysis method, and the most important parameters are extracted. In the process of dimension reduction, firstly, all energy consumption data are subjected to standardization processing, and differences caused by different metering units are eliminated; then, inputting the energy consumption data of the combined label into a dimension reduction model, calculating a weighting kernel matrix, a characteristic value and a characteristic vector, and unitizing the characteristic vector; the cumulative contribution rate is set to 0.85, and the feature is finally reduced to two dimensions according to the value of the cumulative contribution rate.
And then carrying out sufficient incremental learning training on the data after dimensionality reduction. First round of incremental learning is performed: calculating the similarity value between the first group of data, taking an average similarity value alpha, then calculating the similarity value between the newly added data and the first group of data, and increasing the weight value of the existing sample when the similarity value is greater than the existing average similarity value; and when the similarity value is smaller than the existing average similarity value, reducing the weight value of the existing sample and storing all the data together for the incremental learning of the next round. In the next round of incremental learning: firstly grouping the existing data, then calculating the similarity value between the first group of data, calculating to obtain an average similarity value alpha, then calculating the similarity values between the other newly added data and the first group of data, and increasing the weight value of the existing sample when the similarity value is greater than the average similarity value; and when the similarity value is smaller than the average similarity value, reducing the weight value of the existing sample and storing all data together for carrying out the next round of incremental learning, wherein the incremental learning is finished until 1/2 that the average similarity value of the data for incremental learning at a certain time is smaller than the initial average similarity value, and at the moment, deleting the data with the ownership weight value smaller than the initial weight value 1/2 to finish the whole process of sufficient incremental learning. Wherein, the average similarity value alpha, all the energy consumption data and the corresponding weight values of each round are continuously and dynamically adjusted. Through training with sufficient incremental learning, the resulting data distribution is shown in table 2:
TABLE 2 data distribution
Figure BDA0002600827230000131
Figure BDA0002600827230000141
712 data obtained by sufficient incremental learning in table 2 are then input into an error back propagation model (BP) for energy consumption prediction. Before the energy consumption prediction experiment is carried out, the output dimension and the iteration number of the middle layer need to be selected through the experiment. Firstly, selecting the output dimension of the middle layer, and measuring the energy consumption prediction result by using a Root Mean Square Error (RMSE). All RMSE values and run times are shown in table 3.
TABLE 3 RMSE values and run times
Figure BDA0002600827230000142
As can be seen from table 3, when the middle layer output dimension is 7, the RMSE value is the smallest, 0.000431. The corresponding training RMSE and test RMSE values are shown in fig. 3. The RMSE values for the comparative middle layer output dimension of 4 are shown in fig. 4.
After the middle layer dimension is determined to be 7, the minimum RMSE value can be obtained, and then the iteration number is selected. The number of iterations was set to 3000 and the results obtained are shown in table 4. It can be seen that the difference in RMSE values over the test sample is very small, by comparing the RMSE values for different iterations with the run time, when the number of iterations is 1800, the RMSE value is minimal, 0.000112, and the run time is 691.28 seconds. Wherein, when the number of iterations is 3000, the RMSE values of the training data and the test data are shown in fig. 5. The final selected mid-level output dimension and the number of iterations are 7 and 1800, respectively.
TABLE 4 RMSE values and runtimes
Figure BDA0002600827230000151
Figure BDA0002600827230000161
The data after dimension reduction and sufficient incremental learning are input into the LSTM model and the SVR model for energy consumption prediction, and the obtained RMSE value and the operation time are shown in Table 5.
TABLE 5 RMSE values and run times
Figure BDA0002600827230000162
As can be seen from Table 5, in the improved method, the RMSE value of BP algorithm was minimal, 0.07856 less than that of LSTM, and 0.240259 less than that of SVR. Although the running time is not the shortest, the training time is greatly reduced compared with the LSTM prediction method, and the energy consumption prediction is performed by using the BP algorithm in combination with the reduction and sufficient incremental learning, so that the effect is the best. Compared with the results of energy consumption prediction by using the original BP, original LSTM and original SVR methods, the result of energy consumption prediction by combining the BP algorithm with dimension reduction and sufficient incremental learning is shown in Table 5, and as shown in Table 5, the RMSE value of the method provided by the invention is far smaller than the RMSE value of the energy consumption prediction by using the original method, wherein the RMSE value of the method used by the invention is 0.018945 smaller than that of the original BP, 2.491674 smaller than that of the original LSTM and 1.047864 smaller than that of the original SVR, so that the method of dimension reduction and sufficient incremental learning used by the invention realizes high-efficiency utilization of data. However, from the view of the operation time, because sufficient incremental learning is performed, the time used for energy consumption prediction is longer compared with the energy consumption prediction performed by the original BP method, but the operation time is moderate as a whole, so the combination of the weighted kernel principal component analysis method, the sufficient incremental learning and the BP algorithm provided by the invention is effective for energy consumption prediction, and the requirement of energy consumption prediction of continuous data updating is met.
3. Conclusion
In order to avoid the waste of energy consumption during the production of the hub, the invention combines the methods of dimension reduction, incremental learning and energy consumption prediction, firstly, the dimension reduction is carried out on other parameters while the model information is fully utilized by a weighted kernel principal component analysis method, then, the incremental learning is carried out for multiple times by calculating the similarity value and judging, finally, the BP algorithm is used for predicting the energy consumption, and the RMSE is used as an evaluation index for evaluation, thereby realizing the efficient incremental learning and the accurate energy consumption prediction. From the analysis of experimental results, the data can be used more efficiently through the full increment learning, and meanwhile, the RMSE value of the energy consumption prediction is 0.000112 when the method is combined with the weighted kernel principal component analysis method and the BP algorithm, and is respectively 0.07856, 0.240259, 0.018945, 2.491674 and 1.047864 smaller than the LSTM algorithm combined with the weighted kernel principal component analysis method and the full increment learning, the SVR algorithm combined with the weighted kernel principal component analysis method and the full increment learning, the original BP algorithm directly used, the LSTM algorithm directly used and the SVR algorithm directly used. The method of the invention enables accurate energy consumption prediction of the hub.

Claims (3)

1. An energy consumption prediction method based on sufficient incremental learning is characterized by comprising the following steps:
(1) sorting the multidimensional parameters of the energy consumption data, and extracting the models of different hubs;
(2) reducing the dimension of the sorted energy consumption data by adopting a weighted kernel principal component analysis method;
in the step (2), the weight of the weighted kernel principal component analysis method is calculated by taking the models of different hubs as labels for weight, a weighted kernel matrix is constructed by fusing the information values of the labels in the kernel matrix, and the weight corresponding to the data in the weighted kernel matrix is
Figure FDA0002970654170000011
Wherein u isiRepresenting a weight vector corresponding to the ith sample;
(3) dividing the data subjected to dimensionality reduction into training data and test data, grouping the training data and endowing the grouped training data with an initial weight value;
(4) performing first round learning of incremental learning on the first group of training samples and the weighted values, and calculating to obtain an average similarity value of the first group of training data;
(5) continuously inputting the residual data into an incremental learning model, and respectively comparing the residual data with the first group of training data to judge similarity values to perform different operations;
(6) when the similarity between the samples is greater than the average similarity value, increasing the weight value of the existing samples; when the similarity value between the samples is smaller than the average similarity value, reducing the weight value of the existing samples, uniformly storing the samples together for the incremental learning of the next round, and realizing sufficient incremental learning through multiple rounds of incremental learning;
(7) in the process of the operation, the threshold parameter, the data sample and the weight value of each round of incremental learning are adjusted along with the difference of data;
(8) when the average similarity value of the data used for incremental learning at a time is smaller than 1/2 of the initial average similarity value, the incremental learning is finished, and at the moment, the data with the ownership weight value smaller than the initial weight value 1/2 is deleted;
(9) inputting the obtained data into a BP neural network for energy consumption prediction;
(10) and evaluating the energy consumption prediction result by adopting an evaluation index Root Mean Square Error (RMSE).
2. The method of energy consumption prediction based on full incremental learning of claim 1, wherein: in the step (5), the euclidean distance is used to calculate a similarity value of the multidimensional data, and the similarity value calculation formula is as follows:
Figure FDA0002970654170000021
where dist (x, y) represents the Euclidean distance between two different fault signatures, i.e. the newly added signature x and the existing signature y
Figure FDA0002970654170000022
Wherein x isiAnd yiAnd the ith value represents the new added feature x and the ith value of the existing feature y, the difference between x and y is smaller, and the Euclidean distance is smaller, so that the similarity value is larger, and the similarity between the new added feature and the existing feature is measured.
3. The method of energy consumption prediction based on full incremental learning of claim 1, wherein: the process of sufficient incremental learning in step (6) comprises the following steps:
firstly, giving initial weight values to all data, secondly, calculating similarity values among a first group of data to obtain an average similarity value, then, respectively calculating the similarity values of the other newly added data and the first group of data, and increasing the weight value of the existing sample when the similarity values are larger than the existing average similarity value; when the value is smaller than the existing average similarity value, reducing the weight value of the existing sample and storing all data together for carrying out the increment learning of the next round; in the next round of incremental learning, firstly grouping the existing data, secondly calculating the similarity value between the first group of data, calculating to obtain an average similarity value, then respectively calculating the similarity values of the other newly added data and the first group of data, and increasing the weight value of the existing sample when the similarity value is greater than the average similarity value; when the average similarity value is smaller than the average similarity value, the weight value of the existing sample is reduced, all data are stored together for incremental learning in the next round, until 1/2 that the average similarity value of the data for incremental learning is smaller than the initial average similarity value at a certain time, the incremental learning is finished, at this time, the data with the ownership weight value smaller than the initial weight value 1/2 are deleted, and the whole process of sufficient incremental learning is completed.
CN202010723405.3A 2020-07-24 2020-07-24 Energy consumption prediction method based on sufficient incremental learning Active CN111832839B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010723405.3A CN111832839B (en) 2020-07-24 2020-07-24 Energy consumption prediction method based on sufficient incremental learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010723405.3A CN111832839B (en) 2020-07-24 2020-07-24 Energy consumption prediction method based on sufficient incremental learning

Publications (2)

Publication Number Publication Date
CN111832839A CN111832839A (en) 2020-10-27
CN111832839B true CN111832839B (en) 2021-04-30

Family

ID=72924843

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010723405.3A Active CN111832839B (en) 2020-07-24 2020-07-24 Energy consumption prediction method based on sufficient incremental learning

Country Status (1)

Country Link
CN (1) CN111832839B (en)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112613650B (en) * 2020-12-14 2022-06-28 燕山大学 Novel three-stage short-term power load prediction method
CN115032891B (en) * 2022-08-11 2022-11-08 科大智能物联技术股份有限公司 Polycrystalline silicon reduction furnace control method based on time series prediction
CN116996076B (en) * 2023-09-27 2023-12-19 湖北华中电力科技开发有限责任公司 Intelligent management method for electrical energy consumption data of campus equipment

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110901648A (en) * 2018-09-13 2020-03-24 通用汽车环球科技运作有限责任公司 Vehicle, system, and logic for real-time ecological routing and adaptive drive control
CN111260138A (en) * 2020-01-18 2020-06-09 湖南大学 Thermal power generating unit energy consumption dynamic optimization method based on weighting and self-adaptive parallel association rules

Family Cites Families (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR100763551B1 (en) * 2006-11-15 2007-10-04 삼성전자주식회사 Apparatus for reducing contention in prioritized contention access of wireless local network and method using the same
CN104462846B (en) * 2014-12-22 2017-11-10 山东鲁能软件技术有限公司 A kind of equipment fault intelligent diagnosing method based on SVMs
CN104463381B (en) * 2015-01-04 2017-12-08 福州大学 Building energy consumption Forecasting Methodology based on KPCA and WLSSVM
CN106778795A (en) * 2015-11-24 2017-05-31 华为技术有限公司 A kind of sorting technique and device based on incremental learning
CN107316046B (en) * 2017-03-09 2020-08-25 河北工业大学 Fault diagnosis method based on incremental compensation dynamic self-adaptive enhancement
CN107784325B (en) * 2017-10-20 2020-06-23 河北工业大学 Spiral fault diagnosis method based on data-driven incremental fusion
CN110070145B (en) * 2019-04-30 2021-04-27 天津开发区精诺瀚海数据科技有限公司 LSTM hub single-product energy consumption prediction based on incremental clustering

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110901648A (en) * 2018-09-13 2020-03-24 通用汽车环球科技运作有限责任公司 Vehicle, system, and logic for real-time ecological routing and adaptive drive control
CN111260138A (en) * 2020-01-18 2020-06-09 湖南大学 Thermal power generating unit energy consumption dynamic optimization method based on weighting and self-adaptive parallel association rules

Also Published As

Publication number Publication date
CN111832839A (en) 2020-10-27

Similar Documents

Publication Publication Date Title
CN111738512B (en) Short-term power load prediction method based on CNN-IPSO-GRU hybrid model
CN111832839B (en) Energy consumption prediction method based on sufficient incremental learning
CN111027772B (en) Multi-factor short-term load prediction method based on PCA-DBILSTM
CN109461025B (en) Electric energy substitution potential customer prediction method based on machine learning
Cui et al. Research on power load forecasting method based on LSTM model
CN111860982A (en) Wind power plant short-term wind power prediction method based on VMD-FCM-GRU
CN111260136A (en) Building short-term load prediction method based on ARIMA-LSTM combined model
CN113205207A (en) XGboost algorithm-based short-term power consumption load fluctuation prediction method and system
CN102789539B (en) Optimization design method for air vehicle lifting surface structure
CN111861013A (en) Power load prediction method and device
CN113762387B (en) Multi-element load prediction method for data center station based on hybrid model prediction
CN116757057A (en) Air quality prediction method based on PSO-GA-LSTM model
CN113139596A (en) Optimization algorithm of low-voltage transformer area line loss neural network
CN115759415A (en) Power consumption demand prediction method based on LSTM-SVR
CN115099461A (en) Solar radiation prediction method and system based on double-branch feature extraction
CN115995810A (en) Wind power prediction method and system considering weather fluctuation self-adaptive matching
CN115358437A (en) Power supply load prediction method based on convolutional neural network
CN114912195A (en) Aerodynamic sequence optimization method for commercial vehicle
CN110956304A (en) Distributed photovoltaic power generation capacity short-term prediction method based on GA-RBM
CN113033898A (en) Electrical load prediction method and system based on K-means clustering and BI-LSTM neural network
CN112288187A (en) Big data-based electricity sales amount prediction method
CN117151770A (en) Attention mechanism-based LSTM carbon price prediction method and system
CN112330051A (en) Short-term load prediction method based on Kmeans and FR-DBN
Guo et al. Mobile user credit prediction based on lightgbm
CN115796327A (en) Wind power interval prediction method based on VMD (vertical vector decomposition) and IWOA-F-GRU (empirical mode decomposition) -based models

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant